estimation of intensity duration frequency curves for current and future climates
DESCRIPTION
HydrologyTRANSCRIPT
-
1
Estimation of Intensity Duration Frequency Curves
for Current and Future Climates
By Nicolas Desramaut
Department of Civil Engineering and Applied Mechanics McGill University
Montreal, Quebec, Canada
A thesis submitted to the Graduate and Postdoctoral Studies Office in partial fulfilments of requirements of the degree of Master of Engineering
June 2008
Nicolas Desramaut, 2008
All Rights reserved
-
2
Acknowledgement
I would like to kindly and very sincerely thank my supervisor, Professor Van-
Thanh-Van Nguyen, for his never-ending help, his indefectible encouragements,
and his professional guidance.
I would also thank Dr Tan-Danh Nguyen and Sbastien Gagnon for their
assistance in my research work, Gilles Rivard and Pierre Dupuis, for their help
using the softwares, the occupants of the room 391 (Reena, Ali, Arden, Damien,
John, Nabil, Reza, Salman) for their friendship and their permanent support, with
a special mention to Arden, who helped me submitting this thesis despite the
ocean.
And last but not least, I would like to express my deeper thanks to my family, who
despite the distance, were always on my side during my stay in Montreal.
-
i
Abstract
Climate variability and change are expected to have important impacts on the
hydrologic cycle at different temporal and spatial scales In order to build long-
lasting drainage systems, civil engineers and urban planners should take into
account these potential impacts in their hydrological simulations. However, even
if Global Climate Models (GCM) are able to describe the large-scale features of
the climate reasonably well, their coarse spatial and temporal resolutions prevent
their outputs from being used directly in impact assessment models at regional or
local scales.
This study proposes a statistical downscaling approach, based on the scale
invariance concept, to incorporate GCM outputs in the derivation of Intensity-
Duration-Frequency (IDF) curves and the estimation of urban design storms for
current and future climates under different climate change scenarios. The
estimated design storms were then used in the estimations of runoff peaks and
volumes for urban watersheds of different shapes and different levels of surface
imperviousness using the popular Storm Water Management Model (SWMM).
Finally, a regional analysis was performed to estimate the scaling parameters of
extreme rainfall processes for locations with limited or without data. In summary,
results of an illustrative application of the proposed statistical downscaling
approach using rainfall data available in Quebec (Canada) have indicated that it is
feasible to estimate the IDF relations and the resulting design storms and runoff
characteristics for current and future climates in consideration of GCM-based
climate change scenarios. Furthermore, based on the proposed regional analysis of
the scaling properties of extreme rainfalls in Singapore, it has been demonstrated
that it is feasible to estimate the IDF curves for partially-gaged or ungaged sites.
-
ii
Rsum
La variabilit et les changements climatiques devraient avoir des impacts
considrables sur le cycle hydrologique aux diffrentes chelles spatio-
temporelles. Afin de construire des systmes de drainage durables, les ingnieurs
se doivent de prendre en considration ses modifications probables dans leur
simulation. Toutefois, si les Modles de Circulation Globale (MCG) sont capables
de reproduire raisonnablement bien les caractristiques grande chelle du climat,
leurs rsolutions sont trop grossires pour permettre une utilisation immdiate de
leurs informations dans les modlisations urbaines.
Cette tude propose une approche de mise lchelle statistique, base sur les
proprits dinvariance dchelle, afin dincorporer les rsultats des MCG dans la
conception des courbes Intensit-Dure-Frquence (IDF) futures. Ces courbes
sont ensuite utilises pour lestimation de lvolution des quantits de
ruisslement. Enfin, une analyse rgionale permet une valuation des paramtres
de rduction dchelle pour des stations partiellement jauges, voir non jauges.
Cest ainsi que des courbes IDF peuvent tre construites avec un nombre limit de
donnes.
-
iii
Table of Content
Table of Content .................................................................................................................. iii List of Figures ..................................................................................................................... iv 1. Introduction ................................................................................................................. 1
1.1. Context............................................................................................................... 1 1.2. Research Objectives .......................................................................................... 3 1.3. Literature Review ............................................................................................... 4
2. Evaluation of future IDF curves and runoff values in Quebec assuming stationarity in the scaling properties .......................................................................................................... 7
2.1. Data ................................................................................................................... 7 2.2. Methodology .................................................................................................... 13 2.3. Results and Discussion ................................................................................... 20 2.4. Validation ......................................................................................................... 26 2.5. Future Runoff estimations................................................................................ 33 2.6. Discussion ....................................................................................................... 35 2.7. Future researchs .............................................................................................. 35
3. Regional analysis of rainfall scaling properties ........................................................ 36 3.1. Introduction ...................................................................................................... 36 3.2. Data ................................................................................................................. 36 3.3. Methodology .................................................................................................... 43 3.4. Results ............................................................................................................. 44 3.5. Discussion ....................................................................................................... 50
4. Conclusions ............................................................................................................. 51 Bibliography ....................................................................................................................... 52 APPENDICES ................................................................................................................... 56 A - Analysis of the consistency between the different sets of data ............................... 56 B - Computation of GEV Quantile ................................................................................. 57 C - Determination of the three GEV parameters from the 3 first NCM ......................... 58 D - Quantile Plots of the AMP for the baseline period .................................................. 61 E - Evolution of the design storms ................................................................................ 66 F - Evolution of the runoff values in the different watershed configurations ................. 70
-
iv
List of Figures
Figure 1: Localization of the Pierre-Elliot-Trudeau International Airport in Quebec ........ 7 Figure 2: Example of a recorded historical storm ............................................................... 8 Figure 3: Difference between raw SDSM estimates and observed values for the 1961-1994 baseline period ......................................................................................................... 10 Figure 4: Adjustment functions for daily GCM downscaled AMP for the 1961-1994 baseline period .................................................................................................................. 10 Figure 5: Box plots of the 100 adjusted AM Daily Rainfall estimations from CGCM2A2 compared with the observed values (dots), for the 1961-1994 period. ............................. 11 Figure 6: Box plots of the 100 adjusted AM Daily Rainfall estimations from HadCM3A2 compared with the observed values (dots), for the 1961-1994 period. ............................. 11 Figure 7: Difference between observed AMP and recorded historical storms.................. 12 Figure 8: Empirical IDF Curve, baseline period (1961-1994) .......................................... 14 Figure 9: Example of a Desbordes Design Storm ............................................................. 18 Figure 10: Example of a Peyron Design Storm ................................................................ 20 Figure 11: Example of a Watt et al. Design Storm ........................................................... 20 Figure 12: Quantile plots of the 1hr AMP for the baseline period .................................... 21 Figure 13: linearity of the NCM (AMPs, baseline period 1961-1994) ............................. 21 Figure 14: Linearity of the scaling slopes (*k) ............................................................... 22 Figure 15: IDF Curves built with scaling concept for the baseline period symbols= observed values and line = Scaling GEV model ............................................................... 23 Figure 16: three design storms from the observed data, baseline period 1961-1994 ........ 24 Figure 17: Volume estimations from the three design storms based on observed data, baseline period 1961-1994 ................................................................................................ 25 Figure 18: Peak flow estimation, from the three design storms based on observed data, baseline period 1961-1994 ................................................................................................ 25 Figure 19: IDF and design storms conceived from SDSM driven by CGCM2 predictors for 1980-1994, with parameters calibrated over the 1961-1979 period ............................ 26 Figure 20: IDF and design storms conceived from SDSM driven by HadCM3 for 1980-1994, with parameters calibrated over the 1961-1979 period ........................................... 27 Figure 21: Simulated volumes from the Observed GCMs using Watt design storm model vs volume from the historical storms ................................................................................ 27 Figure 22: simulated peak flows from the GCMs using Desbordes design storm models vs volume from the observed storms ..................................................................................... 28 Figure 23: IDF Curves generated from SDSM-CGCM2unbiased daily precipitation, for current and future periods ................................................................................................. 30 Figure 24: Evolution of the daily intensity according to SDSM-CGCM2 with respect to the the baseline period....................................................................................................... 30 Figure 25: IDF Curves generated from HadCM3, for current and future periods ............ 31 Figure 26: Evolution of the daily intensity according to HadCM3 in comparison with the baseline period .................................................................................................................. 31 Figure 27: Design storms for the current and future periods, with a 2-year return period, from both SDSM-GCM. ................................................................................................... 32 Figure 28: Evolution of the runoff peak flows (Watershed, square, 65%, 1ha) ................ 33 Figure 29: Evolution of the runoff volume, (Watershed, square, 65%, 1ha) .................... 34 Figure 30: Spatial repartition of the 9 stations in Singapour ............................................ 38 Figure 31: NCM vs duration to test the scaling behaviour of the AMP in the 9 Singapore stations .............................................................................................................................. 40 Figure 32: Proportionality of the scaling slopes in the 9 Singapore stations .................... 41
-
v
Figure 33: IDF Curves for Ama Keng station derived using Scaling GEV invariance .... 42 Figure 34: daily AM vs. scaling slopes ............................................................................. 44 Figure 35: mean number of rainy days (RD) in January vs. scaling slopes ...................... 45 Figure 36: mean number of rainy days (RD) in February vs. scaling slopes .................... 46 Figure 37: Results of the jackknife method with daily AMP as rainfall parameters for partially gauged site. ......................................................................................................... 47 Figure 38: Results of the jackknife method with number of rainy days in January as rainfall parameters for partially gauged site. ..................................................................... 47 Figure 39: Best result with rainfall parameter =January rainy Day for the Jackknife method for the 9 stations ................................................................................................... 50 Figure 40: Design Storms for the 4 periods for the 2-year return period .......................... 66 Figure 41: Design storms for the 4 periods for the 5-year return period .......................... 67 Figure 42: Design storms for the 4 periods for the 10-year return period ........................ 68 Figure 43: Design storms for the 4 periods for the 50-year return period ........................ 69 Figure 44: Evolution of the runoff peak flows (Rect, 100%, 0.4ha) ................................. 70 Figure 45: Evolution of the runoff peak flows (Rect, 100%, 2 ha) ................................... 71 Figure 46: Evolution of the runoff peak flows (Square, 65%, 10 ha) ............................... 72 Figure 47: Evolution of the runoff volume (Rect, 100%, 0.4 ha) ..................................... 73 Figure 48: Evolution of the runoff volumes (Rect, 100%, 2ha) ........................................ 74 Figure 49: Evolution of the runoff volumes (Square, 65%, 10 ha) ................................... 75
-
vi
List of Tables
Table 1: Repartition of the historical storms during the 1961-1994 baseline period .......... 9 Table 2: Watershed configurations ................................................................................... 16 Table 3: Difference between volumes generated from the observed storms and from the simulated values of design storms .................................................................................... 25 Table 4: Difference between peak flows generated from the observed storms and from the simulated values of design storms .................................................................................... 25 Table 5: Difference (in %) between volumes generated from the observed storms and from the different models of design storms according to respective return periods ......... 28 Table 6: Difference (in %) between peak flows generated from the observed storms and from the different models of design storms according to the respective return periods. .. 28 Table 7: Periods of data availability for daily rainfall and 8 durations for AMP.............. 37 Table 8 : Months of occurrence of the annual maximum daily rainfall ............................ 39 Table 9: Scaling GEV parameters for the 9 Singaporean stations .................................... 42 Table 10: Correlation coefficient between scaling slopes and hydrologic values. ........... 45 Table 11: performance criteria (R) of the jackknife method for partially gauged stations, for the 14 climatic parameters ........................................................................................... 46 Table 12: Performance of the parameter estimation method (com=computed values and obs=observed values, the percentage being the difference between them) ....................... 48 Table 13: performance criteria (R) of the jackknife method for un-gauged stations. ...... 49 Table 14: Relative difference between observed AMP and annual maximum precipitation depths computed from the available historical storms ...................................................... 56
-
1
1. Introduction
1.1. Context
Historically, the first urban drainage systems were built to protect the urban
infrastructures from water damage. Their main task was then to convey
stormwater runoff outside the city as quickly as possible. The opportunity to use
these systems to dispose of wastewater appeared later, in the XIXth century, with
the raising awareness of the impact of polluted water on health. Hence, the
current engineering practice for the design of urban drainage systems requires
taking into account these two main objectives: the removal of wastewater and the
drainage of storm runoff. The first objective could be achieved by considering
different wastewater treatment measures to prevent the discharge of polluted
water into the receiving environment. The second objective is more difficult to
address because of the random variability of runoff in time and in space due to the
randomness of precipitation for current and future climates. Furthermore, the
runoff process could be non-homogeneous due to expanding urbanization of the
watershed causing changes in land use pattern over time (e.g., natural drainage
systems such as streams have been replaced by impervious areas such as streets
and residential buildings). Therefore, in addition to the effects of land use changes
on the runoff process, improved estimations of precipitation patterns for current
and future climates are urgently needed to provide more accurate estimation of
runoff characteristics in the context of a changing climate.
Most urban drainage systems were built to last for a long time, usually more than
50 years, because their repair or retrofitting could be expensive and could cause
massive disruptions of normal city activities since these systems are mainly
located in underground areas. These infrastructures are designed to cope with the
majority (but not all) of the extreme rainfall events that could occur during their
service life (usually ranging from 50 to 100 years). In fact, in view of economic
but also technical constraints, there are some limits to the capacities of drainage
-
2
systems. The current practice has set some criteria for the construction of drainage
systems to balance construction constraints with levels of risk. The minor
drainage system (piped system) should be built to be overwhelmed only every 5 to
10 years, whereas the major drainage system (overland system, including the
roads and parking lots) should support nearly every event (Arisz and Burrell,
2006).
To meet the above-mentioned criteria, engineers need accurate estimations of
potential runoffs from an urban watershed. However, historical runoff data are
generally lacking or have become irrelevant due to changes in land use. In order
to cope with this deficiency, engineers rely on conceptual urban runoff models,
such as the SWMM model, to simulate the rainfall-runoff relations. Indeed, these
models could describe the complex interactions between different hydrologic and
hydraulic processes involved in an urban watershed to transform the input
precipitation pattern into the complete output runoff hydrograph. Hence, to use
these models, knowledge of current and future rainfall patterns, generally
presented as standardized hyetographs, is required. This synthetic pattern of
precipitation provides illustrations of probable distributions of rainfall intensity
over time for a given return period at a local site. In general, these synthetic
hyetographs are computed based on historic rainfall records, with the assumption
that the distributions parameters are stationary or constant over time.
However, recent studies compiled in the Intergovernmental Panel on Climate
Change assessment report (IPCC, 2007) have shown that extreme rainfall events
are very likely to be globally more frequent and more intense in future decades.
Thus, the hypothesis of stationary precipitation occurrences appears highly
questionable in prospect of climate change. Hence, the design storms computed
based on historic data could only be accurate for the current period but not for the
whole service life of drainage systems. Consequently, to build sustainable
drainage systems, engineers need to incorporate future climatic conditions into
-
3
their computation of design storms. The Global Climate Models (GCMs), also
called General Circulation Models, are considered to be able to provide reliable
scenarios for future climatic parameters at the global scale. However, their
resolutions are too coarse (~300 km, 1 day) for being able to be directly used in
the simulation of runoff from urban watersheds since these watersheds are
normally characterized by a short response time (usually, less than 1 day) due to
small surface area size and high imperviousness level. Therefore, outputs from
GCMs should be downscaled to provide climate simulations at appropriate
temporal and spatial scales that are required for accurate estimation of rainfall and
runoff characteristics for urban watersheds.
1.2. Research Objectives
In view of the above-mentioned issues, the present study proposes a new
statistical downscaling method for estimating the Intensity-Duration-Frequency
(IDF) relations that could take into account the climate information given by
GCMs for current as well as for future climatic conditions. More specifically, this
study is aiming at the following particular objectives:
To propose a spatial-temporal statistical downscaling method to describe
the linkage between daily global climate variables and local daily and sub-daily
extreme rainfalls for the derivation of IDF relations for current and future
climates;
To propose a procedure for constructing IDF relations for a local site that
could take into account GCM-based climate change projections;
To propose a procedure for estimating the urban design storms in consideration of
GCM-based climate change scenarios;
To propose a procedure for evaluating the impacts of climate change on
the runoff process for urban watersheds; and to investigate the spatial variability
of extreme rainfall scaling properties over a given region.
-
4
The first part of the study was performed assuming stationary scaling parameters
for extreme rainfalls (i.e., the spatial and temporal scaling properties for future
periods will be the same as the ones found for the current period). The main aim
is to assess the future trends in extreme rainfall events and their consequences on
design rainfalls and urban runoff properties for urban drainage system design.
Notice that the hypothesis of stationary rainfall scaling parameters could only
represent an approximation since these values could change due to the change in
climatic conditions for future periods.
In order to deal with this stationarity issue as well as with the situations where
rainfall records at the site of interest are limited or unavailable, a method is
proposed in the second part to examine the spatial variability of the rainfall
scaling parameters in terms of the regional variation of some climatic variables.
This regional analysis could permit the estimation of IDF curves for current
periods for cases with limited or without data using records from surrounding
stations as well as for the estimation of IDF relations for future periods using
GCM-based projections to avoid the stationarity hypothesis.
1.3. Literature Review
1.3.1. Consequences of Climate Change on Urban Drainage
During the last decades, scientists have become more and more conscious of the
potential impacts of climate change on precipitation patterns for the design of
urban drainage systems. Most previous studies have agreed upon the fact that, as
rainfall events are likely to become more frequent and more intense, the level of
services provided by the drainage systems will be reduced and these urban
structures will suffer more frequent damage, reducing their service life. However,
as Semadeni-Davies (2004) points out, there are neither tools nor guidelines in the
-
5
literature to assess these impacts of climate change in urban catchment areas. In
view of this deficiency, this study proposes the uses of incremental scenarios,
with increases of precipitation arbitrarily taken in a range of the values generated
by the GCM. This method is simple and provides a range of possible
consequences. However, as recognized in the study , these scenarios should not be
considered as future climatic conditions, but only as a spectrum of potential
impacts, as a result of the arbitrary forcing. . Watt et al. (2003) used a similar
approach for estimating the evolution of the runoff values. They assumed that the
rainfall intensity was increased by 15 percent. From the resulting runoff values,
they estimated the parts of the drainage systems that will face overflow capacity.
However, this method is based on crude estimates. They assumed that intensities
of all rainfall events evolved in a similar trend, independently of their duration,
with an increase directly estimated from the GCM large-scale trends. Hence, this
method may not be suitable for urban catchment, where the change in
precipitation patterns might be local but not uniformly proportional.
Denault et al. (2002) proposed an alternative to these methods. They
determined the past linear trends of local rainfall depth for different durations
ranging from 5 min to 1 day. Based on this information, and on the assumption
that the linear increases remain constant, they extrapolated the future intensity-
duration relationships, and, from these data, construct the corresponding design
storms for future periods. They use the software SWMM (Huber and Dickinson,
2002) to perform estimations of future runoff values. This method is an alternative
for assessing potential variations of design storms resulting from climate change,
and therefore for evaluating hydraulic responses of catchment areas, taking into
account that the shape of the design storms can be altered. However, the authors
have to make the hypothesis that the increase rate of the local rainfall depth will
remain the same (i.e. the increase will continue to be linear). The outputs of the
GCM, usually without any persistent linear tendency, make this hypothesis highly
questionable. Moreover, the recent IPCC report (IPCC, 2007) suggests that the
-
6
evolution of anthropogenic activities is very likely to have significant impacts on
future climate conditions. And this evolution would certainly not be linear. So we
need to take into account uncertain nonlinear future causes to assess future effects.
Thus, even if scientists and decision-makers point out the need for a clear
understanding of the effects of climate change on human and social activities,
there is no agreement regarding the suitable approach for evaluating its impacts
on urban hydrologic processes.
1.3.2. Design Storms
Design storms have been used extensively by civil engineers to size sewers for
more than 60 years. With the nonlinear interactions between rainfall and runoff
processes as described by various urban runoff models, these synthetic design
storms are required for the estimation of the complete runoff hydrographs for
urban drainage design purposes. Many frameworks have been conceived in
different countries for the computation of design storms (Marsalek and Watt,
1984), with varying shapes, storm durations, time to peak, maximum intensity and
total volume of rainfall; however, none matches every situation, forcing
hydrologists to perform assessment processes before using a design-storm model
at a new site (Peyron, 2001). Peyron et al. (2005) proposed a procedure to
systematically evaluate design storms models for specific locations. First, the
targeted design storms for different return periods, based on rainfall data from
local Intensity-Duration-Frequency (IDF) curves, were constructed. Then, these
synthetic hyetographs were used as input in the SWMM model to obtain the
respective runoff values (peak flows and volumes). Thus, the runoff properties
estimated from the design storms were compared to those values obtained from
observed historical storms to assess the accuracy of different design storm models.
Based on this approach, the best design storm can be selected for the design of
urban drainage systems.
-
2. Eva
3.1.
This section
present stud
the quality
3.1.1. Dat
For the at-
Trudeau in
Montreal I
verification
future perio
Figure 1
aluation o
statio
Data
n presents a
dy. In addit
as well as th
ta descripti
site analysi
nternational
Island (45.
n and valida
ods, the sim
1: Localizatio
f IDF rela
onarity in
a descriptio
tion, a deta
he suitabilit
ion
is, the stud
airport in
28N, 73.45
ation purpos
mulated (dow
on of the Pier
7
ations and
the rainfa
3.
on of the hy
ailed analysi
ty of the dat
dy site is lo
Dorval, Qu
5W) as sh
ses observe
wnscaled) ra
re-Elliot-Tru
d runoff p
all scaling
ydrologic an
is of these d
ta for this st
ocated at th
uebec (Can
hown in F
ed data are r
ainfall data
udeau Interna
properties
parameter
nd climate d
data is desc
tudy.
he Montrea
nada), in th
Figure 1. F
required. Fo
from GCM
ational Airpo
assuming
rs
data used in
cribed to as
al Pierre-El
he west par
For calibrat
or projectio
s will be us
ort in Quebec
g
n the
ssess
lliot-
rt of
tion,
on in
sed.
c
-
8
3.1.2. Observed data
The observed Annual Maximum Precipitation (AMP) data available at the Pierre
Elliot Trudeau airport for the 1961-1990 period were used in the present study.
The AMP data (representing the maximum rainfall depth fallen each year on a
daily or sub-daily basis) measured in millimetres (within a precision of one-tenth
of millimetres) and available for 9 durations (5, 10, 15, 30 minutes; 1, 2, 6, 12 and
24 hours) were provided by Environment Canada.
Figure 2: Example of a recorded historical storm
To investigate the accuracy of the design storm models, 140 historical rainfall
events (called storms in the remaining of the report ) were recorded during this
1961-1994 baseline period, in millimetres, with a 5-minute time step and with a
one-tenth-millimetre precision. An example of a recorded storm is given in Figure
2. The selection was performed according to their ability to generate the biggest
runoff events and the presence of at least one recorded storm event per year.
0 0 03,1
6,1
36,7
12,2
30,6
12,2
36,7
0
6,1
18,315,3
24,5
39,7
12,2
3,10 0 0 0 0 0
05
1015202530354045
Dep
ths (
mm
)
Time (min)
Historical storm (1961/8/25)
-
9
Year 1961 1962 1963 1964 1965 1966 1967 1968 1969 # of storms 5 2 5 2 2 4 2 1 1
Year 1970 1971 1972 1973 1974 1975 1976 1977 1978 # of storms 1 2 3 5 5 7 5 3 3
Year 1979 1980 1981 1982 1983 1984 1985 1986 1987 # of storms 1 6 3 1 2 4 1 8 8
Year 1988 1989 1990 1991 1992 1993 1994 # of storms 4 5 11 2 3 7 16
Table 1: Repartition of the historical storms during the 1961-1994 baseline period
3.1.3. Simulated data
Daily rainfalls for the baseline 1961-1994 period and for three 30-year future
periods (2010-2039; 2040-2069; and 2070-2099) were provided by Dr. Tan-Danh
Nguyen (Department of Civil Engineering and Applied Mechanics, McGill
University) using the Statistical Downscaling Model (SDSM) (Wilby et al. 2002),
from driven conditions from Global Climate Models (GCM) inputs (Nguyen et
al., 2006). This linear regression based downscaling method allows the
development of the linear statistical linkages between large-scale predictors from
GCMs and local observed predictands (e.g., the daily rainfall in this study). The
GCM climatic variable data are from two different GCMs:
Hadley Centre Coupled Model, version 3 (HadCM3) (Johns et al., 2003), Hadley Centre, United Kingdom
Coupled Global Climate Model, second generation (CGCM2), (Flato and Boer, 2001), Canadian Center for Climate modelling and analysis, Canada
These series of simulations under climate change conditions use the SRES A2
emission scenario with the hypothesis of greenhouse gases (GHG) emissions
following the SRES A2 scenario (Nakicenovic et al., 2000). This scenario is based
on the assumption that the world will become more and more heterogeneous, with
a continuously increasing population (15 billion by 2100), slow and regional
economic growths, and few inter-country technological transfers. This scenario
forecasts that the level of CO2 in the atmosphere will be double in 2100 compared
with the pre-industrial concentration. In the present study, a set of 100 simulations
were performed for each model and for each future period considered.
-
3.1.4. Bia
According
obtained by
the SDSM
correction a
Fi
Nguyen et a
yi
in which yi
AMP estim
The residua
eim
m0, m1 and
Figure 4: Ad
s correctio
to Nguyen
y the statist
procedures
adjustment
igure 3: Diffe
al. (2006) p
yieiyi
i is the adjus
mate and eith
al is comput
m0m1*yi
d m2 are the
djustment fun
n for the d
n et al. (20
tical downsc
and the ob
is required
erence betweefor the 1
propose this
sted daily A
he residual
ted based on
m2*yi e polynomia
nctions for da
10
aily precip
006) , there
caling resul
bserved valu
to improve
en raw SDSM1961-1994 ba
adjustment
AMP in year
associated w
n a second o
al coefficien
aily GCM doperiod.
itation sim
e is a bias
lts from dif
ues, as illus
the precisio
M estimates aseline period
t:
r i, yi the co
with yi.
order regres
nts and the
ownscaled AM
mulation
between th
fferent GCM
strated in . H
on of these p
and observed d
orrespondin
ssion:
e resulting e
MP for the 19
he daily AM
M outputs u
Hence, an b
predictions.
values
ng SDSM d
error.
961-1994 base
MPs
using
bias-
.
daily
eline
-
This adjus
observed an
values (the
Figure 5:
Figure 6: B
stment prov
nnual maxi
lines inside
Box plots of compared wi
Box plots of tcompared wi
vides unbi
imum daily
e the box-pl
the 100 adjusith the observ
the 100 adjusith the observ
11
iased AMP
rainfalls (F
lots) are qui
sted AM Dailved values (d
sted AM Dailved values (d
P estimatio
Figures 5 a
ite similar w
ly Rainfall esdots), for the 1
y Rainfall estdots), for the 1
ons compar
and 6). Inde
with the obs
stimations fro1961-1994 pe
timations fro1961-1994 pe
rable with
eed, the me
erved ones.
om CGCM2Aeriod.
om HadCM3Aeriod.
the
dian
A2
A2
-
12
3.1.5. Selection of years with data inconsistency
In order to compare in a suitable manner the outputs of the computations with the
observed data , the different sets of data have to be consistent. The maximum
depths of precipitation resulting from the available recorded historical storms for
8 durations (5, 10, 15 and 30 minutes; 1, 2, 6 and 12 hours) are computed for each
year, and compared with the corresponding recorded AMP (APPENDICE A).
Three years (1970, 1985 and 1991) have inconsistent sets of data, as illustrated in
Figure 7 with the average of the relative difference over the 8 durations. This
inconsistency is due to the absence of the record of the biggest historical storm in
1970, 1985, and 1991 in our data sets. However, for these 3 years, the different
values were near the median value for each duration. Hence, the removal of these
data should not trigger significant bias in the simulations.
Figure 7: Difference between observed AMP and recorded historical storms
0%
100%
200%
300%
400%
500%
600%
1961
1964
1967
1970
1973
1976
1979
1982
1985
1988
1991
1994
Differen
ce(%
)
Year
-
13
3.2. Methodology
3.2.1. Construction of IDF Curves
3.2.1.1. Generalized extreme values distribution
The General Extreme Values (GEV) distribution is a distribution commonly used
to model extreme events, in particular extreme precipitations. Indeed, it has been
proven to match well the observed extreme values (Nguyen et al., 1998). The
GEV cumulative distribution function can be written as
; , , 1
(1)
where X is the quantile associated with a return period ; , and are
parameters for respectively the location, the scale and the shape.
As a result, the quantile X for the return period can be calculated as follow:
/ ^ (2)
in which p is the probability of exceedance, related to the return period:
/ (3)
3.2.1.2. GEV parameters estimations
The three GEV parameters are estimated by the method of the Non Central
Moments (NCM), where the NCM are defined as:
(4)
It can be demonstrated (Nguyen et al. (2002)) that they are related to the NCM
according to these equations.
(5)
in which (.) is the gamma function
-
14
Thus, using the first three NCMs (k = 1, 2, and 3) given by Equation (4) the three
parameters of the GEV distribution , and can be determined by the method of
moments (see APPENDICE B)
3.2.1.3. Empirical IDF curves
The AMPs were ranked in descending order to compute the exceedance empirical
probability pi for each rank i, using the Cunnanes plotting position formula
(Cunnane, 1978):
.
. (6)
in which n is the number of years of the record.
Then the return periods, assuming that there are no trends in the quantiles values
(i.e., p is invariant over time), are computed as:
(7)
Thus, the empirical frequencies (or return periods) of AMPs can be estimated for
different rainfall durations, and inversely, the AMP quantiles for a given duration
can be computed for specific return periods. The empirical IDF curves can hence
be plotted as shown in Figure 8.
Figure 8: Empirical IDF curves for the 1961-1994 baseline period
-
15
3.2.1.4. Temporal downscaling
There are empirical evidences that some characteristics of short-duration rainfalls
can be deduced from longer duration extreme events (Over and Gupta, 1994,
Aronica and Freni, 2005; and Nhat et al., 2006). Indeed, extreme precipitation
events often present scale invariance properties (Burlando and Rosso, 1996). This
scaling concept can be expressed as:
If a function f is scaling invariant, there is a constant such that: ,
(8)
Nguyen et al. (2002) proved that such functions can be formulated as:
, (9)
Hence, the NCM, for a duration t, can be presented in a similar formulation:
, (10)
(11)
(12)
As a consequence, the three GEV parameters and the quantile X for sub-daily
durations can be computed from daily GEV parameters using these properties
(Nguyen et al., 2006):
(13)
(14)
(15)
(16)
It should be noticed that precipitations can present different scaling regimes. In
other words, they can exhibit scale invariance properties, with different scaling
parameters (i.e. different ) during distinct duration intervals (called scale
-
16
interval). For example, if a function has two scaling regimes, one between t1 and
t2, and the other one between t2 and t3, the function can be described as:
;
;
(17)
Thus, the IDF curves can be fully defined with only the three estimated
parameters for the GEV distribution for the daily duration and the scaling
parameters corresponding to the different extreme rainfall scaling regimes.
3.2.2. Estimation of Runoff Volumes and Peak Flows
3.2.2.1. Watershed Configurations
Different typical urban watersheds have been considered in this part. Their
characteristics are listed in Table 2. They represent some of the configurations
(imperviousness, area, shape) representative of cities in Quebec.
TYPE PERCENT OF IMPERVIOUSNESS(%) AREA (ha) Shape
Parking lots 100 % 0.4 and 2 rectangular
High density residential area 65 % 1 and 10 square
Table 2: Watershed configurations
For illustration purposes, only the results of the small residential area
configuration (square, 65%, 1 ha) are presented in the text and the results for the
other watershed configurations are described in Appendice E
-
17
3.2.2.2. Storm Water Management Model (SWMM)
The Storm Water Management Model (SWMM) developed by the US
Environmental Protection Agency is widely used for the estimation of urban
runoff from rainfall (Huber and Dickinson, 1988). Discrete rainfall hyetographs
are used as input for this model. As the purpose of this study is not to investigate
the performance of the SWMM but to assess the ability of the proposed
downscaling method to mimic actual precipitation pattern and to estimate the
evolution of runoff characteristics due to change in precipitation patterns; hence,
no specific calibration of the SWMM is necessary. Indeed, comparisons between
runoff properties computed from synthetic design storms were compared with
those given by the observed real storms to assess the accuracy of the suggested
estimation method.
In the first step, the runoff values generated by the 136 historical storms were
first computed. Annual maximum volumes and peak flows were extracted from
the results and a frequency analysis of these two extreme series was carried out.
The Gumbel distribution (Extreme Value Type I) provides the best fit for these
runoff value (Mailhot et al., 2007b). Thus the empirical quantiles for the runoff
values can be expressed as:
(18)
. (19)
where is the mean of the runoff volume (or peak flow) and the standard
deviation.
In the second step, synthetic hyetographs for current and future periods were used
to validate the computed runoff for the current period and to make the projection
of runoff for future periods. These synthetic hyetographs are called (synthetic)
design storms.
-
18
3.2.2.3. Design Storms
3.2.2.3.1. General consideration
Peyron (2001) has demonstrated that three design storm models proposed by
Desbordes (1978), Watt et al. (1986), and Peyron et al. (2001) could provide
runoff properties that are similar to those given by the observed historical storms.
These three design storm models were thus chosen for this study. More
specifically, the selected design storms were computed for different return periods
based on the computed IDF curves for Dorval Airport as described previously. In
addition, all these synthetic storms have 1-hour duration with a discrete time step
of 5 minutes.
3.2.2.3.2. Desbordes Design Storm
Figure 9 shows the shape of the Desbordes design storm. The maximum intensity
is located at the 25th minute of the 1-hour storm duration.
Notation:
Equation
;
;
;
;
-
Fig
Fi
Figu
gure 9: Exam
igure 10: Exa
ure 11: Exam
19
mple of a Desb
ample of a Pe
mple of a Wat
bordes Design
eyron Design
tt et al. Desig
n Storm
Storm
gn Storm
-
20
3.2.2.3.3. Peyron Design Storm
The period of maximal intensity lasts 15 minutes between the 20th and 35th
minutes and the intensity peak occurs at the 25th minute (Figure 10). The intensity
is constant outside of the interval. This design storm generates a total depth 1.3
times higher than the actual 1-hour maximal volume for the same return period.
Equation:
. . .
, ; ;
. , ; ; . , ;
3.2.2.3.4. Watt Design Storm
The maximum of intensity takes place in the 5th time interval (Figure 11).
Equation:
, ;
, ;
3.3. Results and Discussion
3.3.1. Verification and Validation Process
In order to assess the feasibility of the proposed downscaling procedure to
generate accurate runoff properties for the current and future periods. The
accuracy of the scaling GEV distribution in the estimation of AMPs for different
durations is evaluated.
3.3.1.1. Accuracy of the Scaling GEV
In order to assert the ability of the GEV distribution to fit observed data, quantile
plots were drawn for each of the 9 durations (Figure 12 and APPENDICES D).
The goodness-of-fit found for all the durations corroborates the hypothesis of
AMPs following GEV distributions.
-
Figu
The scaling
plotting the
the duration
scaling func
So
Thus, if the
proportiona
Figu
ure 12: Quan
g behaviou
e logarithm
n. Indeed,
ction could
ktE
lnkt
e function i
al with a fac
ure 13: Linear
ntile plots of t
ur of the A
of the NCM
as previous
be written
Efk1tk
lnEfk1
s scaling, th
ctor .
rity of the N
21
the 1hr AMP
AMP record
M from the
sly demonst
as
1klnthe plot shou
CM of AMPs
for the basel
ded at the
AMP serie
trated, the n
t
uld be linea
s for the base
line 1961-199
station is i
es, against th
non central
ar and the s
eline 1961-199
94 period
investigated
he logarithm
l moments
(20)
(21)
lopes would
94 period
d by
m of
of a
d be
-
Hence, the
indicates th
are two di
(1=0.50) a
Based on th
correspond
derived fro
the daily A
were comp
Figure 15,
ones (symb
provide acc
e linearity o
he scaling b
ifferent sca
and the seco
Fi
he two estim
ing GEV d
om the thre
AMP. On t
puted, and
the estimat
bols) (R~0
curate descr
of the grap
behaviour of
aling regim
ond between
igure 14: Lin
mated scalin
distribution
ee NCMs an
the basis o
the resultin
ed IDF cur
.99). Hence
ription of th
22
phs exhibit
f the AMPs
mes: the fir
n 30 minute
earity of the
ng paramet
parameters
nd the assoc
of these GE
ng IDF cur
rves (lines)
e, the propo
he AMP dist
ed on Figu
s at Dorval
rst one bet
s and 1 day
scaling slope
ers 1 and
s for the 8
ciated comp
EV distribu
rves were
agreed ve
osed NCM
tributions.
ure 13 and
airport. In p
tween 5 an
y (2=0.21).
s (*k)
2, the three
sub-daily
puted GEV
utions, the
derived. A
ery well wit
M/scaling GE
d on Figure
particular, t
nd 30 min
e NCMs and
durations w
parameters
AMP quan
As indicated
th the empir
EV method
e 14
here
nutes
d the
were
s for
ntiles
d by
rical
can
-
23
Figure 15: IDF Curves built with scaling concept for the baseline period symbols= observed values and line = Scaling GEV model
3.3.1.2. Estimation of runoff properties
The three different design storms (Figure 16) were derived using the computed
IDF curves as described above. These design storms were then used as input to
the SWMM model to estimate the corresponding runoff peaks and volumes for the
selected urban watersheds. Similarly, these two runoff properties were computed
for the historical storms using the SWMM simulation with the same conditions as
for the synthetic design storms.
As expected, the Watt design storm model gave the best performance in the
estimation of the runoff volumes, while the (Figure 17 and Tables 3 and 4)
Desbordes model provided an accurate estimation of the peak flows (Figure 18
and Table 4), and the Peyron model gave accurate estimations for both of these
runoff properties.
-
Figure 16: TThree designn storms from
24 m the observeed data for thhe 1961-1994
4 baseline perriod
-
Tab
Tab
Figure 17: Eand from
Volume Desbordes
Peyron Watt et al. ble 3: Differe
Figure 18: E and from
peak flow Desbordes
Peyron Watt et al. le 4: Differen
Estimation ofm observed st
2 years -17,35% -9,57% 4,21%
ence betweenand f
Estimation ofm observed st
2 years 4,51% 2,46%
35,25% nce between p
and f
25
f runoff volumtorms for the
5 years-16,21%-8,19%4,93%
volumes genfrom the desi
f peak flows torms for the
5 years-3,56%-8,14%23,40%
peak flows gefrom the desi
mes from thr1961-1994 b
10 yea% -15,29%
-7,14%5,59%
nerated from ign storms
from the thr1961-1994 b
10 yea-6,64%-10,01%19,61%
enerated fromign storms
ree design storaseline period
rs 50 ye% -13,4
% -5,23% 6,95
the observed
ee design storaseline period
rs 50 ye% -8,90% -12,3% 14,75m the observe
rms d
ears 40% 3% 5% d storms
rms d
ears 0% 39% 5%
ed storms
-
26
3.4. Validation
The downscaled daily rainfall series were generated using the calibrated SDSM
procedure for two GCMs (the UK HadCM3 and the Canadian CGCM2) for the
1961-1979 period. The design storms and the resulting runoff peak flows and
volumes were then estimated based on the 100 generated downscaled daily
rainfall time series and the bias-correction procedure as suggested in section 2.1.4.
for the 1980-1994 period (Figure 19 and 20). The computed runoff values from
design storms were compared with those from historical storms for the 1980-1994
period.
Figure 19: IDF and design storms estimated based on SDSM downscaling results using CGCM2 predictors for 1980-1994, with parameters calibrated over the 1961-1979 period
Comparable results were found for runoff values computed for these two cases for
both HadCM3 and CGCM2 models (Tables 5 and 6, Figures 21 and 22). The
differences range from 0.5 to 10% for the runoff volumes using Watt model and
between 0.6 and 11.5 % for the peak flows using Desbordes model. Hence, the
-
27
proposed procedure is able to produce consistent IDF curves and to provide good
estimations of runoff values when an appropriate design storm model was
selected.
Figure 20: IDF and design storms conceived from SDSM driven by HadCM3 for 1980-1994, with parameters calibrated over the 1961-1979 period
Figure 21: Simulated runoff volumes from the GCMs using Watt design storm model as compared with the volume from the historical storms (diamond= CGCM2 and square
HadCM3)
0
200
400
600
0 200 400 600
simulated
volum
e(m
3)
observedvolume(m3)
-
28
VOLUMES Return period (years) Desbordes 1,5 2 5 10 20 50 CGCM2 -28,1% -28,1% -26,1% -24,8% -23,6% -22,3% HadCM3 -27,8% -27,2% -24,1% -22,3% -20,8% -19,2%
Peyron 1,5 2 5 10 20 50
CGCM2 -21,9% -21,7% -19,4% -17,9% -16,7% -15,4% HadCM3 -21,5% -20,7% -17,2% -15,2% -13,6% -12,0%
Watt et al. 1,5 2 5 10 20 50 CGCM2 -9,1% -8,5% -6,5% -5,3% -4,3% -3,2% HadCM3 -8,6% -7,3% -4,0% -2,3% -0,9% 0,5%
Table 5: Difference (in %) between volumes generated from the observed storms and from different design storm models for different return periods
Figure 22: Simulated peak flows from the GCMs using Desbordes design storm models as
compared with peakflows from the observed storms(diamond= CGCM2 and square HadCM3)
FLOWS return period (years) Desbordes 1,5 2 5 10 20 50 CGCM2 5,7% -0,8% -7,8% -9,9% -11,1% -11,5% HadCM3 5,7% 0,6% -5,5% -7,3% -7,9% -8,0%
Peyron 1,5 2 5 10 20 50
CGCM2 4,0% -2,8% -11,0% -12,9% -14,0% -14,4% HadCM3 4,0% -1,4% -9,2% -10,3% -10,8% -10,9%
Watt et al. 1,5 2 5 10 20 50 CGCM2 39,2% 30,7% 20,5% 17,4% 15,0% 13,7% HadCM3 39,2% 32,1% 23,7% 20,7% 19,1% 17,9%
Table 6: Difference (in %) between peak flows generated from the observed storms and from different design storm models for different return periods.
0
0,05
0,1
0,15
0,2
0,25
0,3
0,35
0,4
0 0,1 0,2 0,3 0,4
simulated
flow
s(m
3/s)
observedflows(m3/s)
-
29
3.4.1. Simulations for future periods
3.4.1.1. IDF Curves for future periods
As part of the reliability of the results depends on the length of the calibration
period, the calibration steps take into account all the observed data (i.e. the
calibration period will be the baseline periods, 1961-1994, without the 3 years
with inconsistent data). IDF curves, design storms and runoff estimations are
produced for the baseline period using SDSM-GCM unbiased daily simulations
and the three future periods (2020s, 2050s and 2080s), for both GCM (CGCM2
and HadCM3).
The GCM provides two different trends for future extreme rainfalls (Figures 24
and 26). Even if both predict a small decrease of intensities, for all return periods,
during the first period (2020s), the downscaled runs using the CGCM2
predictors show a significant increase (~ +10%) for the last 2 periods (2050s,
2080s), whereas the HadCM3 indicates a continuously slight decrease (reaching ~
-4% in 2080s). However, these changes are not identical in proportions for the
different return periods. Indeed, according to the results obtained from CGCM2
outputs (Figures 23 and 24), the decrease during the first 30-year period is more
pronounced for the rarest events (i.e. with the biggest return periods), but the
relative increase in one century would be approximately the same for all the return
period. The data from HadCM3 model provides different results (Figures 25 and
26). The small decrease in intensity will be fairly similar between the different
frequencies of events.
-
Figure 23
Figure 24: E
1intensity(m
m/hr)
: IDF Curves
Evolution of
1
10
00
1
s generated frcurre
the daily intet
10
196119942020s2050s2080s
30
rom SDSM-Cent and futur
ensity accordthe baseline p
100
duratio
CGCM2unbiare periods
ding to SDSMperiod
0
n(min)
CGC
ased daily pre
M-CGCM2 wit
1000
CMA2
ecipitation, fo
th respect to
10000
or
the
-
Figure
Figure 26:
1
10intensity(m
m/hr)
e 25: IDF Cur
Evolution of
1
10
00
1
rves generate
f the daily int
10
19611994
2020s
2050s
2080s
31
ed from HadC
tensity accordbaseline per
100duratio
4
CM3, for cur
ding to HadCriod
0 1on(min)
H
rrent and futu
CM3 in comp
1000
HadCM3A
ure periods
arison with t
10000
the
-
Figure 27:
3.4.1.2.
CGCM
Design storm
Design Sto
M2
ms for the curfrom
32
orms
rrent and futm both SDSM
ture periods, M-GCM.
HadCM3
with a 2-year
3
r return perio
od,
-
33
The design storms are built based on these new curves (e.g. for the 2-year return
period, Figure 27 or Appendices E, Figures 40-43). As expected, the change of
the rainfall intensities are similar to the ones noticed in the IDF curves, with a
slight reduction when the hyetograph are computed based on HadCM3 inputs and
an increase when CGCM2 predictors are used in the downscaling process.
3.5. Future Runoff estimations
SWMM simulations were performed for the different watershed configurations.
Again, the results exhibit different trends according to the GCM selected (Figures
28 and 29). Again, the same conclusions can be drawn: Two different changes are
probable, increase according to CGCM2 (despite the initial decrease) and slight
decrease with HadCM3.
CGCM2 HadCM3
Figure 28: Evolution of the runoff peak flows (Watershed: square, 65%, 1ha)
-
34
CGCM2 HadCM3
Figure 29: Evolution of the runoff volume, (Watershed: square, 65%, 1ha)
Model : Peyron
0
200
400
600
800
2 5 10 50return period (years)
Vol
um
es (
m3
)
1961-1994 2020s 2050s 2080s
Model : Peyron
0
200
400
600
800
2 5 10 50return period (years)
Vol
umes
(m
3)
1961-1994 2020s 2050s 2080s
Model : Watt et al.
0
200
400
600
800
2 5 10 50return period (years)
Vol
umes
(m
3)
1961-1994 2020s 2050s 2080s
HadCM3A2, Model : Watt et al.
0
200
400
600
800
2 5 10 50return period (years)
Vol
umes
(m
3)
1961-1994 2020s 2050s 2080s
-
35
3.6. Discussion
As the different results illustrate, there are contradictories information provided
by the two different GCM, with opposite consequences. According to the
simulations based on the UK HadCM3, runoff values will decrease, so the
drainage systems would not require to be retrofitted. But, the runoff values
generated from the Canadian CGCM2 tend to predict a potential future
overwhelming of the drainage network. So uncertainties remain on the capacity of
the drainage system to carry out runoff water. These uncertainties seem due to the
inherent uncertainties in the downscaling results using different GCMs (i.e. the
main sources of uncertainties are in the scenarios) Hence the proposed procedure
appears to provide reliable estimations for current climate, but no clear signal can
be exploited for future periods due to the inherent uncertainties of the GCMs.
3.7. Future studies
Even if the procedure proposed in the present study seems to be reliable, further
studies are necessary to address the following issues:
Other GCMs and greenhouse gases emission scenarios should be investigated to provide a wider range of probable runoff evolutions.
The potential impacts of climate change on the scaling behaviour of extreme rainfall processes need to be investigated to avoid the assumption
of stationarity in the rainfall scaling parameters for current and future
periods (one possible approach is to find the linkage between the scaling
parameters and the climatic conditions).
-
36
4. Regional analysis of rainfall scaling properties
4.1. Introduction
As indicated in Chapter 2, one of the main advantages of the scaling GEV
approach is its ability to determine the extreme precipitation distributions for
different durations based only on a few parameters; that is, the three parameters
of the GEV distribution for a given duration and the two slopes of the rainfall
scaling functions. However, the estimation of these two slopes requires a certain
amount of rainfall data that may not be sufficiently available or that may not exist
for the site of interest. In order to cope with these issues, this section investigates
the regional variability of the rainfall scaling parameters. In addition, as
mentioned in Section 2.7, this regional analysis could be useful in dealing with the
projections of rainfall scaling parameters for future periods in the context of
climate change.
4.2. Regional analysis
4.2.1. Case study using rainfall data in Singapore
The regional analysis is performed using rainfall records available at nine
raingages located on the Singapore Island (1.20 N 103.50 E, 585 km, Figure 30).
The selection of Singapore location for this case study was based on the more
homogeneous climatic conditions over the whole 585-km2 Singapore area as
compared to the high regional climate variability over the large region of Quebec.
Two different sets of data were available for each station in Singapore: the AMP
for 8 durations (15, 30 and 45 minutes; 1, 3, 6, 12 and 24 hours) and the daily
rainfalls (a day is considered as a rainy day in this study if there is any trace of
rainfall recorded). All the data were given in millimetres, with a precision of 0.1
mm for the daily rainfall and 1mm for the AMP. The periods of data availability
differ between each station (Table 7), so the study is performed on the first 30-
year common period, 1972-2001.
-
37
Stations
Daily Rainfalls Annual Maximum Precipitations
Ama Keng 1961-2007 1960-2007 (1978)2 Changi 1967-2007 1972-2007 Jurong Industrial 1964-2007 1963-2007 Macritche 1961-2007 1960-2007 Paya Lebar 1961-2007 1960-2007 Seletar 1967-2007 1971-2007 Singapore Orchids 1966-2007 1966-2007 (1969, 70, 76) St James 1961-2007 1960-2007 Tengah 1961-2007 1971-2007 Table 7: Periods of data availability for daily rainfalls and for AMPs for eight durations
4.2.2. Data Analysis
4.2.2.1. Monthly repartition of the extreme rainfall occurrences
As expected, the days with the annual maximum daily rainfall depths occur
mostly during the monsoon periods3 (either from November to early February for
the North-East monsoon, or from July to September for the South-East
monsoon). Indeed, as Singapore is located close to the Equator, the Intertropical
Convergence Zone (ITCZ) structure covers the area during the winter month and
favour regular rainfall events (Chia and Foong, 1991). Table 8 shows the regional
variation of the monthly occurrences of AMPs over the whole study region.
2 (years) = data missing for these years
3 Singaporean National Environment Agency, http://app.nea.gov.sg/cms/htdocs/article.asp?pid=1088
-
Figure 30:: Location of
38
f the nine rainngage stationns in Singapoore
-
39
Year Ama Keng Changi Juron Ind MacritchePaya Lebar Seletar
Singapore Orchids St James Tengah
1972 9 9 12 12 12 9 12 12 9 1973 3 2 4 2 2 2 2 3 10 1974 9 2 9 12 4 4 11 9 11 1975 10 9 2 7 6 8 11 9 10 1976 10 12 12 11 7 7 12 10 10 1977 5 11 8 11 2 7 10 11 10 1978 12 12 12 12 12 12 12 12 12 1979 12 10 11 10 10 7 3 10 10 1980 1 1 5 1 1 1 1 8 1 1981 7 12 7 5 12 5 5 7 7 1982 11 12 4 11 8 12 12 4 12 1983 5 8 7 7 9 5 12 8 5 1984 3 2 3 3 2 6 3 3 9 1985 5 12 5 5 7 9 5 9 5 1986 3 12 3 9 12 4 3 9 3 1987 8 1 3 1 1 1 1 1 10 1988 5 9 5 11 11 11 7 5 5 1989 11 11 8 11 11 11 12 11 11 1990 6 5 4 10 12 9 5 5 12 1991 12 12 12 12 12 12 12 8 12 1992 12 11 11 11 11 11 11 11 11 1993 9 3 3 10 10 3 9 8 3 1994 3 11 11 11 12 6 4 3 6 1995 2 1 7 8 2 2 7 7 2 1996 8 2 4 8 3 3 2 9 2 1997 6 8 8 12 12 5 3 1 8 1998 8 12 1 7 12 12 6 1 8 1999 8 12 11 1 11 5 10 5 8 2000 4 1 4 4 3 1 1 11 3
Table 8 : Number of months of occurrence of annual maximum daily rainfalls
4.2.3. The scaling GEV parameters
The same method described in the previous chapter was used to examine the
scaling behaviour of the AMP series at each station. As shown in Figures 31 and
32, the AMPs for every station in Singapore indicate a simple scaling behaviour
with two distinct scaling regimes from 15 minutes to 45 minutes and from 45
minutes to one day. Table 9 provides the values of the slopes for the two distinct
rainfall scaling regimes for all nine stations.
-
40
Figure 31: The scaling behaviour of AMPs for nine Singapore stations
Ama Keng
y = 1016x0.3642
R2 = 0.974
y = 30569x0.5712
R2 = 0.9773
y = 32.513x0.1737
R2 = 0.9698
y = 460.53x1.6427
R2 = 0.9985
y = 59.388x1.0885
R2 = 0.9984
y = 7.716x0.5401
R2 = 0.9983
1.E+00
1.E+01
1.E+02
1.E+03
1.E+04
1.E+05
1.E+06
1.E+07
1 10 100 1000 10000
duration (min)
NC
MChangi
y = 560.62x0.5248
R2 = 0.9949
y = 12614x0.8319
R2 = 0.9939
y = 24.595x0.2441
R2 = 0.9948
y = 167.59x1.964
R2 = 0.9921
y = 31.269x1.2806
R2 = 0.9918
y = 5.6022x0.6311
R2 = 0.9918
1.E+00
1.E+01
1.E+02
1.E+03
1.E+04
1.E+05
1.E+06
1.E+07
1 10 100 1000 10000
duration (min)
NC
M
Jurond Industrial
y = 986.18x0.3921
R2 = 0.9768
y = 31505x0.6009
R2 = 0.9757
y = 31.28x0.1914
R2 = 0.978
y = 502.31x1.6658
R2 = 0.9981
y = 62.845x1.1006
R2 = 0.9978
y = 7.8612x0.5471
R2 = 0.9975
1.E+00
1.E+01
1.E+02
1.E+03
1.E+04
1.E+05
1.E+06
1.E+07
1 10 100 1000 10000
duration (min)
NC
M
Macritche
y = 653.72x0.4624
R2 = 0.9872
y = 11690x0.7849
R2 = 0.9882
y = 28.155x0.2061
R2 = 0.9827
y = 371.17x1.6888
R2 = 0.9931
y = 49.531x1.1329
R2 = 0.993
y = 6.8947x0.5698
R2 = 0.9928
1.E+00
1.E+01
1.E+02
1.E+03
1.E+04
1.E+05
1.E+06
1.E+07
1 10 100 1000 10000
duration (min)
NC
M
Paya Lebar
y = 562.99x0.5258
R2 = 0.9988
y = 8182x0.9171
R2 = 0.998
y = 26.704x0.2307
R2 = 0.9951
y = 301.1x1.8026
R2 = 0.9951
y = 39.737x1.2263
R2 = 0.9939
y = 6.0479x0.6198
R2 = 0.9927
1.E+00
1.E+01
1.E+02
1.E+03
1.E+04
1.E+05
1.E+06
1.E+07
1 10 100 1000 10000
duration (min)
NC
M
Seletar
y = 846.13x0.4425
R2 = 0.989
y = 14778x0.7939
R2 = 0.9804
y = 33.01x0.188
R2 = 0.9902
y = 188.83x1.96
R2 = 0.9944
y = 31.795x1.3073
R2 = 0.9941
y = 5.5644x0.6527
R2 = 0.9938
1.E+00
1.E+01
1.E+02
1.E+03
1.E+04
1.E+05
1.E+06
1.E+07
1 10 100 1000 10000
duration (min)
NC
M
Singapore Orchids
y = 760.57x0.4428
R2 = 0.9804
y = 13538x0.7778
R2 = 0.9628
y = 30.553x0.1934
R2 = 0.9865
y = 363.51x1.7383
R2 = 0.9972
y = 49.416x1.1576
R2 = 0.9973
y = 6.9444x0.5773
R2 = 0.9974
1.E+00
1.E+01
1.E+02
1.E+03
1.E+04
1.E+05
1.E+06
1.E+07
1 10 100 1000 10000
duration (min)
NC
M
St James
y = 896.48x0.4272
R2 = 0.9881
y = 24779x0.6728
R2 = 0.992
y = 30.809x0.2028
R2 = 0.9813
y = 434.84x1.7155
R2 = 0.9979
y = 56.286x1.1391
R2 = 0.9976
y = 7.4174x0.5681
R2 = 0.9971
1.E+00
1.E+01
1.E+02
1.E+03
1.E+04
1.E+05
1.E+06
1.E+07
1 10 100 1000 10000
duration (min)
NC
M
Tengah
y = 950.96x0.4192
R2 = 0.9818
y = 25111x0.6865
R2 = 0.9815
y = 32.192x0.1927
R2 = 0.9803
y = 283.08x1.8342
R2 = 0.999
y = 40.956x1.2272
R2 = 0.9984
y = 6.2553x0.6148
R2 = 0.9978
1.E+00
1.E+01
1.E+02
1.E+03
1.E+04
1.E+05
1.E+06
1.E+07
1 10 100 1000 10000
duration (min)
NC
M
-
41
Figure 32: Linearity of the slopes of the rainfall scaling behaviour for nine Singapore raingages
Ama Keng
R2 = 1
R2 = 0.99940
0.20.40.60.8
11.21.41.61.8
1 2 3order
beta
Changi
R2 = 0.9998
R2 = 0.9993
0
0.5
1
1.5
2
2.5
1 2 3order
beta
Jurond Industrial
R2 = 1
R2 = 0.9999
00.20.40.60.8
11.21.41.61.8
1 2 3order
beta
Macritche
R2 = 1
R2 = 0.9958
00.20.40.60.8
11.21.41.61.8
1 2 3order
beta
Paya Lebar
R2 = 0.9998
R2 = 0.9935
0
0.5
1
1.5
2
1 2 3order
beta
Seletar
R2 = 1
R2 = 0.99150
0.5
1
1.5
2
2.5
1 2 3order
beta
Singapore Orchids
R2 = 1
R2 = 0.9929
0
0.5
1
1.5
2
1 2 3order
beta
St James
R2 = 1
R2 = 0.9993
0
0.5
1
1.5
2
1 2 3order
beta
Tengah
R2 = 1
R2 = 0.9977
0
0.5
1
1.5
2
1 2 3order
beta
-
4.2.4. IDF
The IDF c
approach a
distribution
other durati
two identif
rainfall dur
parameters
estimated fr
Figure 3
Table 9: S
F Curves
curves for n
as for Dorv
n for the da
ions were th
fied rainfal
rations can b
and 2 scal
from the rain
33: IDF Curv
Ama KengChangi Jurong IndMacritchePaya LebarSeletar Spore OrchSt James Tengah
Scaling GEV p
nine Singap
val station
ily AMP w
hen comput
ll scaling r
be estimated
ling slopes)
nfall quantil
ves for Ama K
42
be0.0.
ustrial 0.0.
r 0.0.
hids 0.0.0.
parameters f
pore station
in Quebec
were first est
ted using th
regimes. He
d using only
). Finally, t
les given by
Keng station d
eta 1 beta 544 0.182642 0.261551 0.196564 0.233611 0.266653 0.225578 0.22557 0.214613 0.21
for nine Sing
ns were co
. The three
timated, an
he scaling fu
ence, the G
y five param
the IDF cur
y these estim
derived using
2 2 1 6 3 6 5 54
gapore statio
onstructed u
e parameter
nd the GEV
function par
GEV distri
meters (3 da
rves for eac
mated GEV
g Scaling GEV
ns
using the s
rs of the G
parameters
rameters for
butions for
aily AMP G
ch station w
V distribution
V invariance
same
GEV
s for
r the
r all
GEV
were
ns.
e
-
43
4.3. Methodology
4.3.1. Assessment of relationships between the scaling function slopes and climatic data
As suggested by Yu et al. (2004), the scaling parameters of AMPs might be
related to the number of rainy days at a local site. In order to investigate this
linkage, the correlation coefficients between the scaling slopes and the rainfall
parameters (either the daily AMPs, the mean number of rainy days per year or per
month) were computed. The correlation coefficient is defined as:
,
(22)
in which E is the expected value and the standard deviation. The relationships
are also graphically assessed with plots of the scaling slopes against each of the
14 climatic parameters (average number of rainy days for each month, for the
whole year and average AMP for each station for the whole period).
4.3.2. Estimation of GEV parameters for a partially gauged station
In this case, neither the sub-daily AMP nor the scaling slopes but the 14 rainfall
parameters are known for the study site. The surrounding stations have all these
data. A simple linear regression is established from the surrounding stations,
between both scaling slopes and each of the 14 rainfall parameters. Then, using
the corresponding parameter at the partially gauged site, the scaling slopes at this
site are estimated. The performance of this linear approximation of the scaling
slopes is evaluated using the jackknife method; that is, the scaling slopes at each
site are supposed missing, then estimated from the other eight surrounding
stations, and finally compared with the empirical one at the site, hence, generating
9x2 couples (empirical/estimated) for both scaling slopes.
-
44
The performance criterion used is the R criterion:
1
(23)
In which is the empirical value of the scaling slope, is the
corresponding estimated value, n the number of stations (here, n=9) and
the variance of the series of the empirical values of scaling slopes.
4.3.3. Estimation of GEV parameters for an ungauged station
In this case, it is assumed that rainfall records (i.e., scaling slopes and rainfall
parameters) are not available at the site of interest. The procedure is similar to the
one used for a partially gauged station, except that the rainfall parameters are
estimated from the ones recorded at the other stations using the formula:
,
,
,
,
(24)
in which is the parameter to be estimated at station i, the corresponding
parameter recorded at station j and , the distance between the stations i and j.
4.4. Results
4.4.1. Relationship between scaling slopes and rainfall parameters
The correlation coefficients, computed for every rainfall parameters are shown in
Table 10 and some typical results of the linear regressions are displayed on
Figures 34, 35 and 36. The scaling slopes are more heavily correlated with the
daily AMP and the number of rainy days in January and February. These results
could be explained by the fact that these 2 months correspond to the North-East
monsoon season, which generates the more intense storms (Chia and Foong,
1991).
-
So if there
will be an i
inversely, t
are high, th
duration AM
Fig
Table 1
Figure 34
are more ra
intense stor
the scaling
he scaling
MPs.
gure 35: Mea
10: Correlatio
4: Relation be
ainy days in
rm this year
slopes will
slopes are m
an number of
daily AM
annual January
FebruaryMarch April May June July
August September
October NovemberDecember
on coefficient
45
etween daily
n one of the
r, so shorter
l decrease. I
more likely
f rainy days (R
beta10.72-0.08-0.61-0.51-0.38-0.180.000.060.250.28-0.11-0.29-0.10-0.10
ts between sca
AM and the
ese months,
r duration A
In the other
y to be high
RD) in Janua
beta20.89-0.33-0.83-0.76-0.62-0.54-0.34-0.19-0.040.12-0.01-0.61-0.24-0.26
aling slopes a
scaling slope
it is more l
AMP will in
r hand, if th
h too, to in
ary vs. scaling
2 9 3 3 6 2 4 4 9 4 2 1 1 4 6 and hydrolog
es
likely that t
ncrease, an
he daily AM
ndicate sho
g slopes
gic variables.
there
d so
MPs
orter
-
Fig
4.4.2. Esti
As can be s
estimates fo
gives also c
resulting co
too differen
consequenc
Table 11: P
ure 36: Mean
imation of
seen from t
or both scal
coherent ap
omputed sca
nt compared
ce, the scalin
Performance c
n number of r
scaling slop
the correlati
ling slopes (
pproximate
aling param
d with the ob
ng slopes ca
daily AM
annual JanuaryFebruaryMarch April May June July
AugustSeptember
OctoberNovemberDecember
criterion (Rfor the
46
rainy days (R
pes for par
ion coeffici
(1 and 2).
for the high
meters obtain
bserved one
an be estim
beta1M 0.37
-0.72-0.01-0.15-0.47-0.55-0.67-0.55-0.45-0.66
r -0.42-0.51
r -0.72r -0.56) of the jackke 14 climatic p
RD) in Febru
rtially gaug
ents, the da
The numbe
her duration
ned with the
es (Table 11
mated fairly w
beta20.70-0.470.480.30-0.19-0.27-0.41-0.54-0.46-0.34-0.38-0.18-0.69-0.51
knife methodparameters.
ary vs. scalin
ged stations
aily AMP pr
er of rainy d
n scaling re
e jackknife
1, Figures 3
well with th
2
7
9 7 1 4 6 4 8 8 9 1 for partially
ng slopes
s
rovides the
days in Janu
egime (2).
method are
37 and 38). A
his procedur
y gauged stati
best
uary
The
e not
As a
re.
ions
-
47
Figure 37: Results of the jackknife method with daily AMP as rainfall parameters for partially gauged site.
Figure 38: Results of the jackknife method with number of rainy days in January as rainfall
parameters for partially gauged site.
-
48
4.4.3. Estimation of scaling slopes for un-gauged stations
obs com AM obs com wRD Obs com MayRD obs com OctRD Ama Keng 111.6 127.2 13.98% 49.5 48.3 -2.31% 16.1 16.9 4.70% 18.2 18.0 -0.68% Changi 145.9 134.0 -8.14% 31.3 43.5 38.73% 13.3 15.3 15.26% 15.3 15.9 4.11% Jurond Ind 125.2 125.5 0.30% 35.6 47.6 33.85% 14.4 16.4 13.51% 16.8 17.4 3.80% Macritche 125.3 130.6 4.27% 45.7 43.6 -4.57% 15.7 15.6 -0.58% 17.2 16.4 -4.83% Paya Lebar 140.3 133.0 -5.21% 41.9 41.1 -1.95% 14.5 15.4 5.69% 15.1 16.4 8.97% Seletar 130.9 131.1 0.13% 44.2 44.1 -0.19% 16.9 15.6 -7.36% 16.7 16.5 -0.82% Spore Orchad 127.0 126.4 -0.48% 49.8 44.7 -10.15% 18.0 15.9 -11.61% 17.6 17.1 -2.99% St James 130.8 128.1 -2.10% 43.1 43.5 0.79% 13.8 15.7 13.62% 14.7 16.9 14.98% Tengah 127.0 116.4 -8.35% 50.0 47.1 -5.67% 17.2 16.0 -7.06% 18.4 17.8 -3.43%
obs comp aRD obs comp JanRD Obs comp JunRD obs comp NovRD Ama Keng 189.8 193.1 1.78% 15.8 15.9 0.66% 13.6 13.7 1.01% 20.6 20.9 1.38% Changi 150.4 176.9 17.65% 12.4 13.6 10.23% 13.0 13.3 1.76% 19.0 20.1 5.88% Jurond Ind 159.1 187.7 17.95% 15.8 15.3 -3.19% 13.2 13.6 3.31% 19.8 20.5 3.50% Macritche 180.5 176.9 -1.97% 14.5 14.4 -1.02% 14.0 13.3 -4.43% 20.7 20.0 -3.76% Paya Lebar 173.3 172.4 -0.54% 12.7 13.9 9.26% 12.9 13.5 4.94% 19.9 19.9 -0.04% Seletar 183.4 178.8 -2.49% 13.7 14.3 4.24% 13.9 13.5 -3.38% 20.3 20.3 -0.13% Spore Orchad 196.5 180.8 -7.98% 15.8 14.9 -5.69% 14.0 13.6 -3.34% 21.0 20.4 -3.25% St James 161.3 177.9 10.30% 13.9 14.7 5.48% 12.5 13.6 9.15% 18.2 20.4 11.85% Tengah 198.1 184.9 -6.65% 16.1 15.6 -3.22% 13.7 13.5 -1.49% 21.1 20.5 -2.92% obs com sprRD obs com FebRD Obs com JulRD obs com DecRD Ama Keng 49.4 51.5 4.31% 13.3 13.3 -0.23% 13.4 13.5 0.74% 20.4 20.1 -1.38% Changi 36.0 44.1 22.81% 9.7 11.1 14.89% 12.8 13.2 2.83% 19.3 19.2 -0.51% Jurond Ind 41.9 48.8 16.61% 12.6 12.7 1.15% 12.6 13.4 6.53% 18.8 19.9 5.88% Macritche 44.9 45.2 0.68% 11.9 11.8 -0.88% 13.6 13.2 -3.39% 19.3 19.3 -0.01% Paya Lebar 42.1 43.4 3.06% 10.2 11.3 10.88% 13.0 13.3 2.90% 19.0 19.3 1.28% Seletar 47.5 45.2 -4.83% 11.5 11.8 2.58% 13.9 13.3 -4.68% 19.0 19.5 2.52% Spore Orchad 51.9 46.6 -10.21% 13.5 12.2 -9.64% 13.8 13.4 -3.31% 20.5 19.5 -5.01% St James 39.3 45.4 15.44% 10.9 12.1 10.24% 12.1 13.4 10.03% 18.3 19.4 6.29% Tengah 53.3 48.1 -9.72% 13.5 13.0 -3.59% 13.6 13.3 -2.03% 20.3 20.1 -1.21% obs com sumRD obs com MarRD Obs comp AugRD Ama Keng 40.9 41.6 1.58% 16.1 15.9 -0.79% 14.0 14.3 2.61% Changi 39.1 40.8 4.30% 12.6 13.9 10.38% 13.8 14.3 3.61% Jurond Ind 40.2 41.4 2.89% 15.2 15.3 0.75% 13.8 14.3 4.22% Macritche 42.0 40.9 -2.60% 14.8 14.3 -3.82% 14.4 14.4 -0.42% Paya Lebar 40.0 41.2 2.91% 13.2 14.0 6.43% 14.1 14.4 2.04% Seletar 43.1 41.2 -4.40% 14.2 14.5 1.65% 15.2 14.4 -5.01% Spore Orchad 43.5 41.3 -5.18% 16.0 14.9 -6.48% 15.7 14.3 -8.95% St James 37.7 41.4 9.81% 12.3 14.8 20.92% 13.1 14.4 9.54% Tengah 41.6 41.0 -1.34% 16.2 15.7 -3.04% 14.3 14.1 -1.13% obs comp falRD obs comp AprRD Obs comp SepRD Ama Keng 53.7 53.4 -0.69% 18.8 18.3 -2.68% 14.9 14.9 -0.27% Changi 43.3 50.5 16.73% 14.9 16.1 7.83% 14.2 14.8 4.23% Jurond Ind 44.8 52.8 17.96% 16.2 17.7 9.41% 14.2 14.9 5.25% Macritche 53.0 50.1 -5.49% 16.6 16.6 -0.33% 15.1 14.7 -2.24% Paya Lebar 49.7 49.3 -0.97% 15.3 16.4 7.09% 14.7 14.6 -0.66% Seletar 51.6 51.1 -1.01% 17.2 16.6 -3.02% 14.6 15.0 2.67% Spore Orchad 54.9 51.2 -6.80% 18.6 17.2 -7.66% 16.3 14.7 -9.62% St James 46.7 50.8 8.85% 14.4 16.9 17.33% 13.7 14.9 8.57% Tengah 54.4 52.3 -3.75% 18.7 18.2 -2.77% 14.9 14.9 0.30%
Table 12: Performance of the parameter estimation method (com=computed values and obs=observed values, the percentage being the difference between these values)
-
49
The proposed formula (Equation 24) for estimating rainfall parameters at an
ungaged site gives results comparable to the observed values (Table 12). The
stations with large discrepancies (Changi and Jurong Industrial Waterworks) are
the ones with a large distance between them. In addition, their geographical
situations near the coast of the island could explain these discrepancies as well
since they have fewer neighbours with similar coastal effects, they have lower
probability of having other stations with similar hydrologic features. Nevertheless,
even theses estimates are fairly similar with the observed values.
On the basis of these good estimates, the slope of the second scaling regime is
fairly well predicted at ungaged stations when the linear regression is based on the
number of rainy days either in January or in February (Table 13 and Figure 39).
However, the method is not able to do better than a random process to evaluate 1 (R
-
Figure 39: R
4.5.
The presen
slopes and
study have
slopes base
interest has
local inform
(over than
improve th
without rai
estimation
al., 2004) th
Results of the
Discussion
nt study ai
some rain
indicated th
ed on infor
s some info
mation, only
45 minute
he estimatio
infall recor
methods or
hat could pr
Jacknife met
n
ims at inve
nfall param
hat it is po
rmation fro
ormation re
y the slope o
es) can be
on of the
ds further
r to find oth
rovide a stro
50
thod based onthe 9 statio
estigating th
meters using
ossible to es
om surround
egarding the
of the temp
reasonably
scaling fun
studies are
her rainfall
onger relatio
n the numberons
he relations
g a regiona
stimate fair
ding station
e rainfall
oral downs
y estimate
nctions for
needed to
l parameters
on with the
r of rainy day
ships betwe
l analysis.
rly well the
ns as long
parameters.
caling for lo
ed. Howeve
location w
o consider o
s (see, for
scaling slop
ys in January
een the sca
Results of
rainfall sca
as the sit
. If there is
onger durat
er, in orde
with limited
other nonlin
instance, Y
pes.
y for
aling
this
aling
e of
s no
tions
er to
d or
near
Yu et
-
51
5. Conclusions
The main objective of this stud